Self-supervised Learning of Adversarial Example: Towards Good Generalizations for Deepfake Detection

This repository contains PyTorch implementation of the CVPR oral presentation paper:

Self-supervised Learning of Adversarial Example: Towards Good Generalizations for Deepfake Detection.

Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, Jue Wang

The proposed method uses adversarial self-supervised training to improve the generability of current deepfake detectors. The pipeline is illustrated in the following figure:

Preparation

pacakges

Please refer to the requirements.txt for details.

pretrained weights

Download Xception pretrained weights and dlib landmark predictor and put them in the weights folder.

Datasets

training datasets

We use the FaceForensicsDataset (FF++) for training. Please go to their project page for downloading. For every video in FF++ dataset, we extract 270 frames for training, and 100 each for evaluation and testing rigously following their data splitting strategy. The data structure is like:

    SLADD project
    |---README.md
    |---...
    |---data
        |---FF
            |---image
                |---FF-DF
                    |---071_054 
                        |---0001.png  
                        |---...
                    |---...
                |---FF-F2F
                |---FF-FS
                |---FF-NT
                |---real
            |---mask
                |---FF-DF
                    |---071_054 
                        |---0001_mask.png  
                        |---...
                    |---...
                |---FF-F2F
                |---FF-FS
                |---FF-NT
            |---config
                |---train.json
                |---test.json
                |---eval.json

test datasets

We use the DFDC, CelebDF, and DF1.0 for testing. These datasets are organized similar to FF++. Please go to their sites for downloading.

Running

python train.py  --resolution 256 --dataname none --dset FF-DF --meta FF-DF -n 1 -g 8 -nr 0  -mp 5555

Citation

If you find this code useful for your research, please cite:

  @inproceedings{chen2022self,
      author = {Liang Chen and Yong Zhang and Yibing Song and Lingqiao Liu and Jue Wang},
      title = {Self-supervised Learning of Adversarial Examples: Towards Good Generalizations for DeepFake Detections}, 
      booktitle = {CVPR},  
      year = {2022}
  }

Contact

Please open an issue or contact Liang Chen (liangchen527@gmail.com) if you have any questions or any feedback.

liangchen527 / SLADD

readme